A new research paper explores the phenomenon of Large Language Models (LLMs) generating incorrect answers despite having the correct information accessible within their intermediate processing states. The study, titled "On the Tip of the Tongue: Why LLMs Hallucinate Answers They Can Decode," differentiates between the model's ability to 'read' the correct token from its internal states and its 'write' capability to select that token as the final output. Researchers found that even when the correct answer is readable, LLMs often fail to select it, attributing this to the selection margin at the final readout stage. Modifying the answer support to levels typical of successful generations can recover first-token selection for many failures, though this does not always lead to a fully correct answer. AI
IMPACT Provides insight into the internal mechanisms of LLM failures, potentially guiding future research on improving model accuracy and reliability.
RANK_REASON Research paper published on arXiv detailing a specific aspect of LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
